image-evaluator

image-evaluator is a lightweight CLI utility for generative AI practitioners and researchers. It evaluates six core quality dimensions of synthetic images: predicted visual aesthetics, text-to-image semantic alignment, facial identity preservation, and pairwise image fidelity (deep perceptual distance LPIPS, structural similarity SSIM, and peak signal-to-noise ratio PSNR).

Evaluation Workflow

flowchart LR
    A[Task Goal] --> B{Choose --metrics}
    B -->|aesthetic| C[LAION Aesthetic]
    B -->|clip| D[CLIP Similarity]
    B -->|arcface| E[ArcFace Distance]
    B -->|lpips| F[LPIPS Distance]
    B -->|ssim| G[SSIM Similarity]
    B -->|psnr| H[PSNR Ratio]
    C --> I[Float, Higher Better]
    D --> J[Cosine Sim, Higher Better]
    E --> K[Cosine Dist, Lower Better]
    F --> L[Distance, Lower Better]
    G --> M[Index, Higher Better]
    H --> N[dB, Higher Better]

Metric Selection

Quick Start

Installation

0.1.0a1 is an alpha preview and requires Python 3.11–3.14. Install the preview explicitly because package installers normally exclude prereleases:

python -m pip install --pre image-evaluator==0.1.0a1

The supported runtime path is macOS or Linux with CPU ONNX Runtime. Linux users who want the GPU runtime can replace onnxruntime with onnxruntime-gpu after installation. Windows is currently unverified.

Tutorial

The CLI enforces explicit metric selection via --metrics and initializes only selected models:

• --metrics (required): One or more of aesthetic, clip, arcface, lpips, ssim, psnr.
• --image (required): Path to an image file or directory.
• --prompt: Required when clip is selected; prohibited otherwise.
• --reference: Required when reference-based metrics (arcface, lpips, ssim, psnr) are selected; prohibited otherwise.

1. Aesthetic evaluation only:

   image-evaluator --metrics aesthetic --image path/to/image.png
   

2. CLIP text alignment only:

   image-evaluator --metrics clip --image path/to/image.png --prompt "a cat in oil painting style"
   

3. Pairwise fidelity triad (LPIPS, SSIM, PSNR):

   image-evaluator --metrics lpips ssim psnr --image path/to/image.png --reference path/to/ref.png
   

4. Multi-metric evaluation across folders:

   image-evaluator --metrics aesthetic clip arcface lpips ssim psnr \
       --image path/to/images/ \
       --prompt path/to/prompts/ \
       --reference path/to/refs/
   

Interpretation & Protocol Guidelines

1. Protocol Consistency: Always compare scores under identical model backbones and preprocessing pipelines.
2. Relative Comparison: Avoid universal absolute thresholds; interpret scores relative to a baseline control.
3. Fail-Fast Spatial Dimension Policy: Pairwise metrics (lpips, ssim, psnr) strictly reject mismatched image dimensions with ValueError to prevent artificial interpolation distortion. Align sizes beforehand via downsampling or super-resolution.
4. SSIM Minimum Size: SSIM requires both image dimensions to be at least 11 pixels because it uses the documented 11 × 11 Gaussian window. Smaller inputs fail with ValueError.

Preview Status

See CHANGELOG.md for the accepted user-facing changes and known preview limitations.